리샘플링: SMOTE와 언더샘플링

One way to help a model see the rare class is to resample: change how many examples of each class it trains on.

You can add more of the minority (oversample) or trim the majority (undersample). Both aim to even out the class counts.

The simplest oversample just copies minority rows until counts match. It works, but those exact duplicates can make a model memorize them.

SMOTE creates synthetic minority examples instead of copies. It invents new, plausible points between real minority neighbors.

SMOTE picks a minority row, finds a nearby minority neighbor, and places a fresh point somewhere on the line between them.

The imbalanced-learn library makes SMOTE a two-line affair on your training features and labels.

Undersampling drops majority rows until balance returns. It trains faster but throws away data, which can cost real signal.

Oversample when data is scarce and every row matters. Undersample when the majority is huge and you can spare some rows for speed.

Resample only the training set. Touch the test set and your scores become fantasy, since real data is never rebalanced for you.

SMOTE must learn from training rows alone. Leaking test rows into resampling inflates results and quietly causes data leakage. ⚠️

Balanced counts help, but they are not magic. Pair resampling with the right metrics and sometimes class weights for the best results.

What makes SMOTE different from plain oversampling?

Resampling balances the training set by oversampling (SMOTE makes synthetic rows) or undersampling. Never resample the test set. 🎯